#50 🎂 Paris Women in Machine Learning — AI and Male Fantasies of Self-Engendering and Low-Rank Optimisation
On October 11, 2024, we celebrated our 50th meetup, a significant milestone in our ongoing mission to support and promote women and gender minorities in the fields of machine learning and data science. This special edition was hosted by Pruna.ai in the heart of Paris’s 9th arrondissement.
Chloé and Caroline, two of our co-founders, kicked off the event with a reflective look at our community’s journey. Founded in 2017 with just 60 participants, WiMLDS Paris has since blossomed into a dynamic and thriving community, and now boasts over 5,500 members on Meetup and 3,500 LinkedIn followers, making it the second-largest chapter globally!
Over the years, we have hosted 139 speakers, featuring a mix of first-time presenters and renowned experts. We have actively collaborated with other tech groups through initiatives such as 14 paper reading sessions and joint meetups, fostering a rich exchange of ideas and knowledge within the broader tech community!
The first talk of the evening was delivered by Isabelle Collet, Associate Professor at the University of Geneva. She delved into the historical and philosophical underpinnings of artificial intelligence, focusing on the male-dominated narratives surrounding creation and control. Drawing on the works of early AI pioneers like Norbert Wiener and Alan Turing, she discussed how the origins of AI are deeply tied to the idea of self-engendering — the ability to create intelligent, self-replicating systems without human reproduction.
Isabelle examined how this concept, rooted in myths like the Golem and popularised in works such as Mary Shelley’s Frankenstein, has shaped the development of AI. She highlighted that early AI research often focused on activities perceived as highly intellectual but also traditionally masculine, such as proving theorems or playing chess. This narrow focus on male-coded tasks, like chess, overshadowed other forms of intelligence, such as language processing or emotional recognition, which may have been considered more representative of human cognition if developed by a more gender-diverse group of scientists.
Isabelle also explored the notion of “creating in one’s own image,” a theme prevalent in the development of machines that mimic human intelligence. She argued that this desire to create autonomous, intelligent machines reflects deeper societal anxieties about control and reproduction. In particular, she pointed out how some narratives frame the act of creation as a replacement for women’s role in reproduction, a recurring theme in both science fiction and AI development. Ultimately, Isabelle called for a broader and more inclusive perspective in AI, one that recognises the contributions of women and challenges the biases present in the field.
If you want to know more, check out her slides below:
Then, Irène Waldspurger, a CNRS researcher at Paris-Dauphine University, gave an insightful talk on low-rank optimisation. Low-rank optimization are problems where we want to minimize a function f over a space of matrices, given a constraint on the rank of these matrices. One of the most well-known applications of this method is the Netflix problem, where algorithms predict user preferences by identifying patterns in sparse data using low-rank models.
She explained that low-rank optimisation is particularly useful when the data being analysed has inherent structure, allowing for this simplification. She introduced two primary approaches for solving these problems: convexified algorithms and non-convex algorithms. Convexified algorithms work by approximating the original non-convex problem with a convex one, which is easier to solve using standard optimization techniques. These methods are more straightforward, often relying on tools like the nuclear norm to encourage low-rank solutions.
On the other hand, non-convex algorithms approach the problem directly, without simplifying it. Although they are more computationally challenging, she highlighted that non-convex algorithms can be very effective, especially in scenarios where the data has a lot of randomness or redundancy. These methods can sometimes avoid local minima and reach optimal solutions even in complex settings.
Irène emphasised the ongoing research challenges in this field, particularly in finding efficient ways to handle non-convex problems and improving the speed and accuracy of convexified approaches. She concluded by discussing the importance of rigorous theoretical guarantees for these algorithms, ensuring they not only perform well in practice but also have a strong mathematical foundation to back their effectiveness in real-world applications.
If you want to know more, check out her slides below:
We’re gearing up for an exciting event in December. Stay tuned for more updates as we get closer to the date!
If you do not want to miss our events, you can:
đź”— follow us on Twitter, Meetup, and LinkedIn
đź“‘ check our Google spreadsheet if you want to speak đź“Ł, host đź’™, help đźŚ
đź“Ťjoin our Slack channel for information, discussions, and opportunities
đź“© send an email to the Paris WiMLDS team to paris@wimlds.org
🎬 subscribe to our WiMLDS Paris medium page and Youtube channel
📸 follow the global WiMLDS on Instagram, LinkedIn, and Facebook
🔥 share your company or lab’s job offers for free on the global WiMLDS’ website.
